Increasing the Transferability of Adversarial Attack by Depthwise Separation Convolution

Deep neural networks, susceptible to adversarial examples, can be misled by almost imperceptible alterations. In a black-box setting, however, adversarial examples created using the source model typically exhibit limited transferability for attacking other defended models. Therefore, obtaining adversarial examples with high transferability is highly desirable as it facilitates practical adversarial attacks. In our research, we introduce a novel approach termed DSC-Attack, designed to create adversarial examples with enhanced transferability that effectively bypass defense models. Differing from conventional attack algorithms, we employ depthwise separable convolutions to enhance computational efficiency, conserve computing resources, thereby achieving more transferable malicious adversaries. Experimental results on the widely-recognized ImageN et dataset show that although DSC-Attack under the single model setting performs consistently with existing attack methods, it significantly outperforms current methods in terms of transferability within an ensemble-model framework.